Experience
Software Engineer II
Microsoft, Outlook Copilot 2022 – Present- Owned LoRA SFT post-training for time-expression understanding, hitting 95% accuracy on the golden set while cutting the system prompt by 80% — the org's first proof that post-training generalizes here, laying the groundwork for the current post-training roadmap
- Architected an iterative rejection-sampling fine-tuning pipeline that scores live LLM rollouts and promotes top performers into the training set, enabling preference optimization on a single-digit GPU LoRA + SFT pipeline
- Collaborated on RL post-training pipeline design and implementation, launching experimental training runs on 100+ GPU clusters and reaching 82% pass rate on internal calendar reasoning benchmarks
- Developed a self-correcting synthetic-data pipeline using LLM correction loops to generate a full synthetic org spanning 50+ users and 100K+ Outlook calendar entities, powering training and eval across 5+ features
- Designed an evaluation pipeline for Event Importance classification, a core LLM ranking signal feeding 6+ downstream features, driving it to 75% positive user feedback in production
- Built core agentic-harness infrastructure powering Outlook Copilot — spanning containerized agent execution, persistent storage, and calendar integration — serving 1.36M+ weekly active users, a 40x increase from 32K
- Engineered skill orchestration and tool usage for the meeting scheduler agent, routing skill selection, tool calls, and calendar parsing through a single reliable agent loop and lifting kept meeting suggestions 5x in production
- Created an LLM constraint parser and solver for scheduling that translates natural-language preferences into formal Z3 constraints, reaching 95% scheduling accuracy versus 30% for the existing rule-based baseline
Research Assistant
University of Michigan, Nallasamy Lab 2020 – 2022- Developed a surgical tool identification model with a lightweight CNN and post-process prediction averaging, achieving state-of-the-art performance with 30% fewer parameters
- Built a custom 3D DenseNet by stacking filter weights in the temporal dimension for spatio-temporal feature extraction, achieving state-of-the-art results on cataract surgery step identification
Publications
Nicholas Matton, et al. "Analysis of Cataract Surgery Instrument Identification Performance of Convolutional and Recurrent Neural Network Ensembles Leveraging BigCat"
Translational Vision Science and Technology, Volume 11, Issue 4, 2022
Ossama Mahmoud, Han Zhang, Nicholas Matton, et al. "CatStep: Automated Cataract Surgical Phase Classification and Boundary Segmentation Leveraging Inflated 3D-Convolutional Neural Network Architectures and BigCat"
Ophthalmology Science, Volume 4, Issue 1, 2024
Education
M.S. Computer Science and Engineering
University of Michigan, Ann Arbor 2017 – 20223.92/4.00 GPA
B.S. Computer Science and Engineering
University of Michigan, Ann Arbor 2017 – 20223.92/4.00 GPA, Minor in Mathematics, Summa Cum Laude
